IC-182Truthfulness encoding in Mistral-7B, Mistral-7B-Instruct, Llama3-8B, and Llama3-8B-Instruct is skill-specific rather than universal; probing classifiers do not meaningfully generalize across different task types beyond logit-based baselines
Hadas Orgad, Michael Toker, Zorik Gekhman, Roi Reichart, Idan Szpektor, Hadas Kotek, Yonatan Belinkov
The paper trains probing classifiers on one dataset and tests them on nine others, covering factual retrieval, common-sense reasoning, bias, arithmetic, and sentiment. While raw AUC values above 0.5 suggest some cross-task generalization, subtracting the performance of the strongest logit-based baseline (logit-min-exact) reveals that most of the apparent generalization is already captured by output logits alone. Meaningful probe-specific generalization is limited to tasks requiring similar skills (e.g., TriviaQA to HotpotQA, Winobias to Winogrande). This challenges the hypothesis of a single universal truthfulness direction in LLM representations.
Evidence
correlational
Caveat
The generalization results are presented in heatmaps (Figures 3, 7, 8, 9) rather than tabulated numbers. The paper notes some unexplained patterns, such as asymmetric generalization from TriviaQA to Math tasks.